Search results for: Imbalanced dataset

Authors: Sara Rejeb, Catherine Duveau, Tabea Rebafka

Abstract:

To monitor the production process of turbofan aircraft engines, multiple measurements of various geometrical parameters are systematically recorded on manufactured parts. Engine parts are subject to extremely high standards as they can impact the performance of the engine. Therefore, it is essential to analyze these databases to better understand the influence of the different parameters on the engine's performance. Self-organizing maps are unsupervised neural networks which achieve two tasks simultaneously: they visualize high-dimensional data by projection onto a 2-dimensional map and provide clustering of the data. This technique has become very popular for data exploration since it provides easily interpretable results and a meaningful global view of the data. As such, self-organizing maps are usually applied to aircraft engine condition monitoring. As databases in this field are huge and complex, they naturally contain multiple missing entries for various reasons. The classical Kohonen algorithm to compute self-organizing maps is conceived for complete data only. A naive approach to deal with partially observed data consists in deleting items or variables with missing entries. However, this requires a sufficient number of complete individuals to be fairly representative of the population; otherwise, deletion leads to a considerable loss of information. Moreover, deletion can also induce bias in the analysis results. Alternatively, one can first apply a common imputation method to create a complete dataset and then apply the Kohonen algorithm. However, the choice of the imputation method may have a strong impact on the resulting self-organizing map. Our approach is to address simultaneously the two problems of computing a self-organizing map and imputing missing values, as these tasks are not independent. In this work, we propose an extension of self-organizing maps for partially observed data, referred to as missSOM. First, we introduce a criterion to be optimized, that aims at defining simultaneously the best self-organizing map and the best imputations for the missing entries. As such, missSOM is also an imputation method for missing values. To minimize the criterion, we propose an iterative algorithm that alternates the learning of a self-organizing map and the imputation of missing values. Moreover, we develop an accelerated version of the algorithm by entwining the iterations of the Kohonen algorithm with the updates of the imputed values. This method is efficiently implemented in R and will soon be released on CRAN. Compared to the standard Kohonen algorithm, it does not come with any additional cost in terms of computing time. Numerical experiments illustrate that missSOM performs well in terms of both clustering and imputation compared to the state of the art. In particular, it turns out that missSOM is robust to the missingness mechanism, which is in contrast to many imputation methods that are appropriate for only a single mechanism. This is an important property of missSOM as, in practice, the missingness mechanism is often unknown. An application to measurements on one type of part is also provided and shows the practical interest of missSOM.

Keywords: imputation method of missing data, partially observed data, robustness to missingness mechanism, self-organizing maps

Procedia PDF Downloads 126

Authors: Farnaz Alborzi, Angelika Schmitz, Rainer Stamminger

Abstract:

Reducing water temperatures in the wash phase of a washing programme and increasing the overall cycle durations are the latest trends in decreasing energy consumption of washing programmes. Since the implementation of the new energy efficiency classes in 2010, manufacturers seem to apply the aforementioned washing strategy with lower temperatures combined with longer programme durations extensively to realise energy-savings needed to meet the requirements of the highest energy efficiency class possible. A semi-representative on-line survey in eleven European countries (Czech Republic, Finland, France, Germany, Hungary, Italy, Poland, Romania, Spain, Sweden and the United Kingdom) was conducted by Bonn University in 2015 to shed light on consumer opinion and behaviour regarding the effects of the lower washing temperature and longer cycle duration in laundry washing on consumers’ acceptance of the programme. The risk of the long wash cycle is that consumers might not use the energy efficient Standard programmes and will think of this option as inconvenient and therefore switch to shorter, but more energy consuming programmes. Furthermore, washing in a lower temperature may lead to the problem of cross-contamination. Washing behaviour of over 5,000 households was studied in this survey to provide support and guidance for manufacturers and policy designers. Qualified households were chosen following a predefined quota: -Involvement in laundry washing: substantial, -Distribution of gender: more than 50 % female , -Selected age groups: -20–39 years, -40–59 years, -60–74 years, -Household size: 1, 2, 3, 4 and more than 4 people. Furthermore, Eurostat data for each country were used to calculate the population distribution in the respective age class and household size as quotas for the consumer survey distribution in each country. Before starting the analyses, the validity of each dataset was controlled with the aid of control questions. After excluding the outlier data, the number of the panel diminished from 5,100 to 4,843. The primary outcome of the study is European consumers are willing to save water and energy in a laundry washing but reluctant to use long programme cycles since they don’t believe that the long cycles could be energy-saving. However, the results of our survey don’t confirm that there is a relation between frequency of using Standard cotton (Eco) or Energy-saving programmes and the duration of the programmes. It might be explained by the fact that the majority of washing programmes used by consumers do not take so long, perhaps consumers just choose some additional time reduction option when selecting those programmes and this finding might be changed if the Energy-saving programmes take longer. Therefore, it may be assumed that introducing the programme duration as a new measure on a revised energy label would strongly influence the consumer at the point of sale. Furthermore, results of the survey confirm that consumers are more willing to use lower temperature programmes in order to save energy than accepting longer programme cycles and majority of them accept deviation from the nominal temperature of the programme as long as the results are good.

Keywords: duration, energy-saving, standard programmes, washing temperature

Procedia PDF Downloads 199

Authors: Noura Al Akkari, Aurélie Foucquier, Sylvain Lespinats

Abstract:

Load monitoring is a management process for energy consumption towards energy savings and energy efficiency. Non Intrusive Load Monitoring (NILM) is one method of load monitoring used for disaggregation purposes. NILM is a technique for identifying individual appliances based on the analysis of the whole residence data retrieved from the main power meter of the house. Our NILM framework starts with data acquisition, followed by data preprocessing, then event detection, feature extraction, then general appliance modeling and identification at the final stage. The event detection stage is a core component of NILM process since event detection techniques lead to the extraction of appliance features. Appliance features are required for the accurate identification of the household devices. In this research work, we aim at developing a new event detection methodology with accurate load disaggregation to extract appliance features. Time-domain features extracted are used for tuning general appliance models for appliance identification and classification steps. We use unsupervised algorithms such as Dynamic Time Warping (DTW). The proposed method relies on detecting areas of operation of each residential appliance based on the power demand. Then, detecting the time at which each selected appliance changes its states. In order to fit with practical existing smart meters capabilities, we work on low sampling data with a frequency of (1/60) Hz. The data is simulated on Load Profile Generator software (LPG), which was not previously taken into consideration for NILM purposes in the literature. LPG is a numerical software that uses behaviour simulation of people inside the house to generate residential energy consumption data. The proposed event detection method targets low consumption loads that are difficult to detect. Also, it facilitates the extraction of specific features used for general appliance modeling. In addition to this, the identification process includes unsupervised techniques such as DTW. To our best knowledge, there exist few unsupervised techniques employed with low sampling data in comparison to the many supervised techniques used for such cases. We extract a power interval at which falls the operation of the selected appliance along with a time vector for the values delimiting the state transitions of the appliance. After this, appliance signatures are formed from extracted power, geometrical and statistical features. Afterwards, those formed signatures are used to tune general model types for appliances identification using unsupervised algorithms. This method is evaluated using both simulated data on LPG and real-time Reference Energy Disaggregation Dataset (REDD). For that, we compute performance metrics using confusion matrix based metrics, considering accuracy, precision, recall and error-rate. The performance analysis of our methodology is then compared with other detection techniques previously used in the literature review, such as detection techniques based on statistical variations and abrupt changes (Variance Sliding Window and Cumulative Sum).

Keywords: electrical disaggregation, DTW, general appliance modeling, event detection

Procedia PDF Downloads 45

Authors: Muhammad Ahnaf Zahin, Yaw Adu-Gyamfi

Abstract:

Bus transit is a crucial component of transportation networks, especially in urban areas. Any intelligent transportation system must have accurate real-time information on bus travel times since it minimizes waiting times for passengers at different stations along a route, improves service reliability, and significantly optimizes travel patterns. Bus agencies must enhance the quality of their information service to serve their passengers better and draw in more travelers since people waiting at bus stops are frequently anxious about when the bus will arrive at their starting point and when it will reach their destination. For solving this issue, different models have been developed for predicting bus travel times recently, but most of them are focused on smaller road networks due to their relatively subpar performance in high-density urban areas on a vast network. This paper develops a deep learning-based architecture using a single-step multi-station forecasting approach to predict average bus travel times for numerous routes, stops, and trips on a large-scale network using heterogeneous bus transit data collected from the GTFS database. Over one week, data was gathered from multiple bus routes in Saint Louis, Missouri. In this study, Gated Recurrent Unit (GRU) neural network was followed to predict the mean vehicle travel times for different hours of the day for multiple stations along multiple routes. Historical time steps and prediction horizon were set up to 5 and 1, respectively, which means that five hours of historical average travel time data were used to predict average travel time for the following hour. The spatial and temporal information and the historical average travel times were captured from the dataset for model input parameters. As adjacency matrices for the spatial input parameters, the station distances and sequence numbers were used, and the time of day (hour) was considered for the temporal inputs. Other inputs, including volatility information such as standard deviation and variance of journey durations, were also included in the model to make it more robust. The model's performance was evaluated based on a metric called mean absolute percentage error (MAPE). The observed prediction errors for various routes, trips, and stations remained consistent throughout the day. The results showed that the developed model could predict travel times more accurately during peak traffic hours, having a MAPE of around 14%, and performed less accurately during the latter part of the day. In the context of a complicated transportation network in high-density urban areas, the model showed its applicability for real-time travel time prediction of public transportation and ensured the high quality of the predictions generated by the model.

Keywords: gated recurrent unit, mean absolute percentage error, single-step forecasting, travel time prediction.

Procedia PDF Downloads 47

Authors: R. Tamborrino, F. Rinaudo

Abstract:

Digital technologies offer possibilities to effectively convert historical archives into instruments of knowledge able to provide a guide for the interpretation of historical phenomena. Digital conversion and management of those documents allow the possibility to add other sources in a unique and coherent model that permits the intersection of different data able to open new interpretations and understandings. Urban history uses, among other sources, the inventories that register human activities in a specific space (e.g. cadastres, censuses, etc.). The geographic localisation of that information inside cartographic supports allows for the comprehension and visualisation of specific relationships between different historical realities registering both the urban space and the peoples living there. These links that merge the different nature of data and documentation through a new organisation of the information can suggest a new interpretation of other related events. In all these kinds of analysis, the use of GIS platforms today represents the most appropriate answer. The design of the related databases is the key to realise the ad-hoc instrument to facilitate the analysis and the intersection of data of different origins. Moreover, GIS has become the digital platform where it is possible to add other kinds of data visualisation. This research deals with the industrial development of Turin at the beginning of the 20th century. A census of factories realized just prior to WWI provides the opportunity to test the potentialities of GIS platforms for the analysis of urban landscape modifications during the first industrial development of the town. The inventory includes data about location, activities, and people. GIS is shaped in a creative way linking different sources and digital systems aiming to create a new type of platform conceived as an interface integrating different kinds of data visualisation. The data processing allows linking this information to an urban space, and also visualising the growth of the city at that time. The sources, related to the urban landscape development in that period, are of a different nature. The emerging necessity to build, enlarge, modify and join different buildings to boost the industrial activities, according to their fast development, is recorded by different official permissions delivered by the municipality and now stored in the Historical Archive of the Municipality of Turin. Those documents, which are reports and drawings, contain numerous data on the buildings themselves, including the block where the plot is located, the district, and the people involved such as the owner, the investor, and the engineer or architect designing the industrial building. All these collected data offer the possibility to firstly re-build the process of change of the urban landscape by using GIS and 3D modelling technologies thanks to the access to the drawings (2D plans, sections and elevations) that show the previous and the planned situation. Furthermore, they access information for different queries of the linked dataset that could be useful for different research and targets such as economics, biographical, architectural, or demographical. By superimposing a layer of the present city, the past meets to the present-industrial heritage, and people meet urban history.

Keywords: digital urban history, census, digitalisation, GIS, modelling, digital humanities

Procedia PDF Downloads 167

Authors: Lee Tasker, Ali Karrech, Jeffrey Shragge, Matthew Josh

Abstract:

Monitoring for the deterioration of concrete infrastructure is an important assessment tool for an engineer and difficulties can be experienced with monitoring for deterioration within an infrastructure. If a failure crack, or fluid seepage through such a crack, is observed from the surface often the source location of the deterioration is not known. Geophysical methods are used to assist engineers with assessing the subsurface conditions of materials. Techniques such as Ground Penetrating Radar (GPR) provide information on the location of buried infrastructure such as pipes and conduits, positions of reinforcements within concrete blocks, and regions of voids/cavities behind tunnel lining. This experiment underlines the application of GPR as an infrastructure-monitoring tool to highlight and monitor regions of possible deterioration within a concrete test wall due to an increase in the generation of fractures; in particular, during a time period of applied load to a concrete wall up to and including structural failure. A three-point load was applied to a concrete test wall of dimensions 1700 x 600 x 300 mm³ in increments of 10 kN, until the wall structurally failed at 107.6 kN. At each increment of applied load, the load was kept constant and the wall was scanned using GPR along profile lines across the wall surface. The measured radar amplitude responses of the GPR profiles, at each applied load interval, were reconstructed into depth-slice grids and presented at fixed depth-slice intervals. The corresponding depth-slices were subtracted from each data set to compare the radar amplitude response between datasets and monitor for changes in the radar amplitude response. At lower values of applied load (i.e., 0-60 kN), few changes were observed in the difference of radar amplitude responses between data sets. At higher values of applied load (i.e., 100 kN), closer to structural failure, larger differences in radar amplitude response between data sets were highlighted in the GPR data; up to 300% increase in radar amplitude response at some locations between the 0 kN and 100 kN radar datasets. Distinct regions were observed in the 100 kN difference dataset (i.e., 100 kN-0 kN) close to the location of the final failure crack. The key regions observed were a conical feature located between approximately 3.0-12.0 cm depth from surface and a vertical linear feature located approximately 12.1-21.0 cm depth from surface. These key regions have been interpreted as locations exhibiting an increased change in pore-space due to increased mechanical loading, or locations displaying an increase in volume of micro-cracks, or locations showing the development of a larger macro-crack. The experiment showed that GPR is a useful geophysical monitoring tool to assist engineers with highlighting and monitoring regions of large changes of radar amplitude response that may be associated with locations of significant internal structural change (e.g. crack development). GPR is a non-destructive technique that is fast to deploy in a production setting. GPR can assist with reducing risk and costs in future infrastructure maintenance programs by highlighting and monitoring locations within the structure exhibiting large changes in radar amplitude over calendar-time.

Keywords: 4D GPR, engineering geophysics, ground penetrating radar, infrastructure monitoring

Procedia PDF Downloads 144

Authors: Darren A. Whitaker, Dan Palmer, Jens Wesholowski, James Flaherty, John Mack, Ahmad B. Albadarin, Gavin Walker

Abstract:

Traditional mechanosynthesis has been performed by either ball milling or manual grinding. However, neither of these techniques allow the easy application of process control. The temperature may change unpredictably due to friction in the process. Hence the amount of energy transferred to the reactants is intrinsically non-uniform. Recently, it has been shown that the use of Twin-Screw extrusion (TSE) can overcome these limitations. Additionally, TSE enables a platform for continuous synthesis or manufacturing as it is an open-ended process, with feedstocks at one end and product at the other. Several materials including metal-organic frameworks (MOFs), co-crystals and small organic molecules have been produced mechanochemically using TSE. The described advantages of TSE are offset by drawbacks such as increased process complexity (a large number of process parameters) and variation in feedstock flow impacting on product quality. To handle the above-mentioned drawbacks, this study utilizes UV-Vis spectrophotometry (InSpectroX, ColVisTec) as an online tool to gain real-time information about the quality of the product. Additionally, this is combined with real-time process information in an Advanced Process Control system (PharmaMV, Perceptive Engineering) allowing full supervision and control of the TSE process. Further, by characterizing the dynamic behavior of the TSE, a model predictive controller (MPC) can be employed to ensure the process remains under control when perturbed by external disturbances. Two reactions were studied; a Knoevenagel condensation reaction of barbituric acid and vanillin and, the direct amidation of hydroquinone by ammonium acetate to form N-Acetyl-para-aminophenol (APAP) commonly known as paracetamol. Both reactions could be carried out continuously using TSE, nuclear magnetic resonance (NMR) spectroscopy was used to confirm the percentage conversion of starting materials to product. This information was used to construct partial least squares (PLS) calibration models within the PharmaMV development system, which relates the percent conversion to product to the acquired UV-Vis spectrum. Once this was complete, the model was deployed within the PharmaMV Real-Time System to carry out automated optimization experiments to maximize the percentage conversion based on a set of process parameters in a design of experiments (DoE) style methodology. With the optimum set of process parameters established, a series of PRBS process response tests (i.e. Pseudo-Random Binary Sequences) around the optimum were conducted. The resultant dataset was used to build a statistical model and associated MPC. The controller maximizes product quality whilst ensuring the process remains at the optimum even as disturbances such as raw material variability are introduced into the system. To summarize, a combination of online spectral monitoring and advanced process control was used to develop a robust system for optimization and control of two TSE based mechanosynthetic processes.

Keywords: continuous synthesis, pharmaceutical, spectroscopy, advanced process control

Procedia PDF Downloads 140

Authors: Jianli Jiang, Bai-Chen Xie

Abstract:

The climate issue has aroused global concern. Achieving sustainable development is a good path for countries to mitigate environmental and climatic pressures, although there are many difficulties. The first step towards sustainable development is to evaluate the environmental efficiency of the energy industry with proper methods. The power sector is a major source of CO2, SO2, and NOx emissions. Evaluating the environmental efficiency (EE) of power systems is the premise to alleviate the terrible situation of energy and the environment. Data Envelopment Analysis (DEA) has been widely used in efficiency studies. However, measuring the efficiency of a system (be it a nation, region, sector, or business) is a challenging task. The classic DEA takes the decision-making units (DMUs) as independent, which neglects the interaction between DMUs. While ignoring these inter-regional links may result in a systematic bias in the efficiency analysis; for instance, the renewable power generated in a certain region may benefit the adjacent regions while the SO2 and CO2 emissions act oppositely. This study proposes a spatial network DEA (SNDEA) with a slack measure that can capture the spatial spillover effects of inputs/outputs among DMUs to measure efficiency. This approach is used to study the EE of China's power system, which consists of generation, transmission, and distribution departments, using a panel dataset from 2014 to 2020. In the empirical example, the energy and patent inputs, the undesirable CO2 output, and the renewable energy (RE) power variables are tested for a significant spatial spillover effect. Compared with the classic network DEA, the SNDEA result shows an obvious difference tested by the global Moran' I index. From a dynamic perspective, the EE of the power system experiences a visible surge from 2015, then a sharp downtrend from 2019, which keeps the same trend with the power transmission department. This phenomenon benefits from the market-oriented reform in the Chinese power grid enacted in 2015. The rapid decline in the environmental efficiency of the transmission department in 2020 was mainly due to the Covid-19 epidemic, which hinders economic development seriously. While the EE of the power generation department witnesses a declining trend overall, this is reasonable, taking the RE power into consideration. The installed capacity of RE power in 2020 is 4.40 times that in 2014, while the power generation is 3.97 times; in other words, the power generation per installed capacity shrank. In addition, the consumption cost of renewable power increases rapidly with the increase of RE power generation. These two aspects make the EE of the power generation department show a declining trend. Incorporation of the interactions among inputs/outputs into the DEA model, this paper proposes an efficiency evaluation method on the basis of the DEA framework, which sheds some light on efficiency evaluation in regional studies. Furthermore, the SNDEA model and the spatial DEA concept can be extended to other fields, such as industry, country, and so on.

Keywords: spatial network DEA, environmental efficiency, sustainable development, power system

Procedia PDF Downloads 73

Authors: Noura Al Akkari, Aurélie Foucquier, Sylvain Lespinats

Abstract:

Load monitoring is a management process for energy consumption towards energy savings and energy efficiency. Non Intrusive Load Monitoring (NILM) is one method of load monitoring used for disaggregation purposes. NILM is a technique for identifying individual appliances based on the analysis of the whole residence data retrieved from the main power meter of the house. Our NILM framework starts with data acquisition, followed by data preprocessing, then event detection, feature extraction, then general appliance modeling and identification at the final stage. The event detection stage is a core component of NILM process since event detection techniques lead to the extraction of appliance features. Appliance features are required for the accurate identification of the household devices. In this research work, we aim at developing a new event detection methodology with accurate load disaggregation to extract appliance features. Time-domain features extracted are used for tuning general appliance models for appliance identification and classification steps. We use unsupervised algorithms such as Dynamic Time Warping (DTW). The proposed method relies on detecting areas of operation of each residential appliance based on the power demand. Then, detecting the time at which each selected appliance changes its states. In order to fit with practical existing smart meters capabilities, we work on low sampling data with a frequency of (1/60) Hz. The data is simulated on Load Profile Generator software (LPG), which was not previously taken into consideration for NILM purposes in the literature. LPG is a numerical software that uses behaviour simulation of people inside the house to generate residential energy consumption data. The proposed event detection method targets low consumption loads that are difficult to detect. Also, it facilitates the extraction of specific features used for general appliance modeling. In addition to this, the identification process includes unsupervised techniques such as DTW. To our best knowledge, there exist few unsupervised techniques employed with low sampling data in comparison to the many supervised techniques used for such cases. We extract a power interval at which falls the operation of the selected appliance along with a time vector for the values delimiting the state transitions of the appliance. After this, appliance signatures are formed from extracted power, geometrical and statistical features. Afterwards, those formed signatures are used to tune general model types for appliances identification using unsupervised algorithms. This method is evaluated using both simulated data on LPG and real-time Reference Energy Disaggregation Dataset (REDD). For that, we compute performance metrics using confusion matrix based metrics, considering accuracy, precision, recall and error-rate. The performance analysis of our methodology is then compared with other detection techniques previously used in the literature review, such as detection techniques based on statistical variations and abrupt changes (Variance Sliding Window and Cumulative Sum).

Keywords: general appliance model, non intrusive load monitoring, events detection, unsupervised techniques;

Procedia PDF Downloads 48

Authors: Jia Li, Xing Wei, Yuchen Hong, Yang Lu

Abstract:

Detecting vehicle behavior has always been the focus of intelligent transportation, but with the explosive growth of the number of vehicles and the complexity of the road environment, the vehicle behavior videos captured by traditional surveillance have been unable to satisfy the study of vehicle behavior. The traditional method of manually labeling vehicle behavior is too time-consuming and labor-intensive, but the existing object detection and tracking algorithms have poor practicability and low behavioral location detection rate. This paper proposes a vehicle behavior detection algorithm based on the dual-stream convolution network and the multi-dimensional video dynamic detection network. In the videos, the straight-line behavior of the vehicle will default to the background behavior. The Changing lanes, turning and turning around are set as target behaviors. The purpose of this model is to automatically mark the target behavior of the vehicle from the untrimmed videos. First, the target behavior proposals in the long video are extracted through the dual-stream convolution network. The model uses a dual-stream convolutional network to generate a one-dimensional action score waveform, and then extract segments with scores above a given threshold M into preliminary vehicle behavior proposals. Second, the preliminary proposals are pruned and identified using the multi-dimensional video dynamic detection network. Referring to the hierarchical reinforcement learning, the multi-dimensional network includes a Timer module and a Spacer module, where the Timer module mines time information in the video stream and the Spacer module extracts spatial information in the video frame. The Timer and Spacer module are implemented by Long Short-Term Memory (LSTM) and start from an all-zero hidden state. The Timer module uses the Transformer mechanism to extract timing information from the video stream and extract features by linear mapping and other methods. Finally, the model fuses time information and spatial information and obtains the location and category of the behavior through the softmax layer. This paper uses recall and precision to measure the performance of the model. Extensive experiments show that based on the dataset of this paper, the proposed model has obvious advantages compared with the existing state-of-the-art behavior detection algorithms. When the Time Intersection over Union (TIoU) threshold is 0.5, the Average-Precision (MP) reaches 36.3% (the MP of baselines is 21.5%). In summary, this paper proposes a vehicle behavior detection model based on multi-dimensional dynamic detection network. This paper introduces spatial information and temporal information to extract vehicle behaviors in long videos. Experiments show that the proposed algorithm is advanced and accurate in-vehicle timing behavior detection. In the future, the focus will be on simultaneously detecting the timing behavior of multiple vehicles in complex traffic scenes (such as a busy street) while ensuring accuracy.

Keywords: vehicle behavior detection, convolutional neural network, long short-term memory, deep learning

Procedia PDF Downloads 95

Authors: George Zhou, Yunchan Chen, Candace Chien

Abstract:

Kidney replacement therapy is the current standard of care for end-stage renal diseases. In-center or home hemodialysis remains an integral component of the therapeutic regimen. Arteriovenous fistulas (AVF) make up the vascular circuit through which blood is filtered and returned. Naturally, AVF patency determines whether adequate clearance and filtration can be achieved and directly influences clinical outcomes. Our aim was to build a deep learning model for automated AVF stenosis screening based on the sound of blood flow through the AVF. A total of 311 patients with AVF were enrolled in this study. Blood flow sounds were collected using a digital stethoscope. For each patient, blood flow sounds were collected at 6 different locations along the patient’s AVF. The 6 locations are artery, anastomosis, distal vein, middle vein, proximal vein, and venous arch. A total of 1866 sounds were collected. The blood flow sounds are labeled as “patent” (normal) or “stenotic” (abnormal). The labels are validated from concurrent ultrasound. Our dataset included 1527 “patent” and 339 “stenotic” sounds. We show that blood flow sounds vary significantly along the AVF. For example, the blood flow sound is loudest at the anastomosis site and softest at the cephalic arch. Contextualizing the sound with location metadata significantly improves classification performance. How to encode and incorporate categorical metadata is an active area of research1. Herein, we study ordinal (i.e., integer) encoding schemes. The numerical representation is concatenated to the flattened feature vector. We train a vision transformer (ViT) on spectrogram image representations of the sound and demonstrate that using scalar multiples of our integer encodings improves classification performance. Models are evaluated using a 10-fold cross-validation procedure. The baseline performance of our ViT without any location metadata achieves an AuROC and AuPRC of 0.68 ± 0.05 and 0.28 ± 0.09, respectively. Using the following encodings of Artery:0; Arch: 1; Proximal: 2; Middle: 3; Distal 4: Anastomosis: 5, the ViT achieves an AuROC and AuPRC of 0.69 ± 0.06 and 0.30 ± 0.10, respectively. Using the following encodings of Artery:0; Arch: 10; Proximal: 20; Middle: 30; Distal 40: Anastomosis: 50, the ViT achieves an AuROC and AuPRC of 0.74 ± 0.06 and 0.38 ± 0.10, respectively. Using the following encodings of Artery:0; Arch: 100; Proximal: 200; Middle: 300; Distal 400: Anastomosis: 500, the ViT achieves an AuROC and AuPRC of 0.78 ± 0.06 and 0.43 ± 0.11. respectively. Interestingly, we see that using increasing scalar multiples of our integer encoding scheme (i.e., encoding “venous arch” as 1,10,100) results in progressively improved performance. In theory, the integer values do not matter since we are optimizing the same loss function; the model can learn to increase or decrease the weights associated with location encodings and converge on the same solution. However, in the setting of limited data and computation resources, increasing the importance at initialization either leads to faster convergence or helps the model escape a local minimum.

Keywords: arteriovenous fistula, blood flow sounds, metadata encoding, deep learning

Procedia PDF Downloads 52

Authors: Sachin Jain, Daniel Perez-Astudillo, Dunia A. Bachour, Antonio P. Sanfilippo

Abstract:

The future of solar energy in Qatar is evolving steadily. Hence, high-quality spatial solar radiation data is of the uttermost requirement for any planning and commissioning of solar technology. Generally, two types of solar radiation data are available: satellite data and ground observations. Satellite solar radiation data is developed by the physical and statistical model. Ground data is collected by solar radiation measurement stations. The ground data is of high quality. However, they are limited to distributed point locations with the high cost of installation and maintenance for the ground stations. On the other hand, satellite solar radiation data is continuous and available throughout geographical locations, but they are relatively less accurate than ground data. To utilize the advantage of both data, a product has been developed here which provides spatial continuity and higher accuracy than any of the data alone. The popular satellite databases: National Solar radiation Data Base, NSRDB (PSM V3 model, spatial resolution: 4 km) is chosen here for merging with ground-measured solar radiation measurement in Qatar. The spatial distribution of ground solar radiation measurement stations is comprehensive in Qatar, with a network of 13 ground stations. The monthly average of the daily total Global Horizontal Irradiation (GHI) component from ground and satellite data is used for error analysis. The normalized root means square error (NRMSE) values of 3.31%, 6.53%, and 6.63% for October, November, and December 2019 were observed respectively when comparing in-situ and NSRDB data. The method is based on the Empirical Bayesian Kriging Regression Prediction model available in ArcGIS, ESRI. The workflow of the algorithm is based on the combination of regression and kriging methods. A regression model (OLS, ordinary least square) is fitted between the ground and NSBRD data points. A semi-variogram is fitted into the experimental semi-variogram obtained from the residuals. The kriging residuals obtained after fitting the semi-variogram model were added to NSRBD data predicted values obtained from the regression model to obtain the final predicted values. The NRMSE values obtained after merging are respectively 1.84%, 1.28%, and 1.81% for October, November, and December 2019. One more explanatory variable, that is the ground elevation, has been incorporated in the regression and kriging methods to reduce the error and to provide higher spatial resolution (30 m). The final GHI maps have been created after merging, and NRMSE values of 1.24%, 1.28%, and 1.28% have been observed for October, November, and December 2019, respectively. The proposed merging method has proven as a highly accurate method. An additional method is also proposed here to generate calibrated maps by using regression and kriging model and further to use the calibrated model to generate solar radiation maps from the explanatory variable only when not enough historical ground data is available for long-term analysis. The NRMSE values obtained after the comparison of the calibrated maps with ground data are 5.60% and 5.31% for November and December 2019 month respectively.

Keywords: global horizontal irradiation, GIS, empirical bayesian kriging regression prediction, NSRDB

Procedia PDF Downloads 62

Authors: M. Regina Carreira-Lopez

Abstract:

Lightweight ontologies seek to concrete union relationships between a parent node, and a secondary node, also called "child node". This logic relation (L) can be formally defined as a triple ontological relation (LO) equivalent to LO in ⟨LN, LE, LC⟩, and where LN represents a finite set of nodes (N); LE is a set of entities (E), each of which represents a relationship between nodes to form a rooted tree of ⟨LN, LE⟩; and LC is a finite set of concepts (C), encoded in a formal language (FL). Mondoc enables more refined searches on semantic and classified facets for retrieving specialized knowledge about Atlantic migrations, from the Declaration of Independence of the United States of America (1776) and to the end of the Spanish Civil War (1939). The model looks forward to increasing documentary relevance by applying an inverse frequency of co-ocurrent hypernymy phenomena for a concrete dataset of textual corpora, with RMySQL package. Mondoc profiles archival utilities implementing SQL programming code, and allows data export to XML schemas, for achieving semantic and faceted analysis of speech by analyzing keywords in context (KWIC). The methodology applies random and unrestricted sampling techniques with RMySQL to verify the resonance phenomena of inverse documentary relevance between the number of co-occurrences of the same term (t) in more than two documents of a set of texts (D). Secondly, the research also evidences co-associations between (t) and their corresponding synonyms and antonyms (synsets) are also inverse. The results from grouping facets or polysemic words with synsets in more than two textual corpora within their syntagmatic context (nouns, verbs, adjectives, etc.) state how to proceed with semantic indexing of hypernymy phenomena for subject-heading lists and for authority lists for documentary and archival purposes. Mondoc contributes to the development of web directories and seems to achieve a proper and more selective search of e-documents (classification ontology). It can also foster on-line catalogs production for semantic authorities, or concepts, through XML schemas, because its applications could be used for implementing data models, by a prior adaptation of the based-ontology to structured meta-languages, such as OWL, RDF (descriptive ontology). Mondoc serves to the classification of concepts and applies a semantic indexing approach of facets. It enables information retrieval, as well as quantitative and qualitative data interpretation. The model reproduces a triple tuple ⟨LN, LE, LT, LCF L, BKF⟩ where LN is a set of entities that connect with other nodes to concrete a rooted tree in ⟨LN, LE⟩. LT specifies a set of terms, and LCF acts as a finite set of concepts, encoded in a formal language, L. Mondoc only resolves partial problems of linguistic ambiguity (in case of synonymy and antonymy), but neither the pragmatic dimension of natural language nor the cognitive perspective is addressed. To achieve this goal, forthcoming programming developments should target at oriented meta-languages with structured documents in XML.

Keywords: hypernymy, information retrieval, lightweight ontology, resonance

Procedia PDF Downloads 101

Authors: Masoud Fathi Kazerouni, Jens Schlemper, Klaus-Dieter Kuhnert

Abstract:

Green planet is one of the Earth’s names which is known as a terrestrial planet and also can be named the fifth largest planet of the solar system as another scientific interpretation. Plants do not have a constant and steady distribution all around the world, and even plant species’ variations are not the same in one specific region. Presence of plants is not only limited to one field like botany; they exist in different fields such as literature and mythology and they hold useful and inestimable historical records. No one can imagine the world without oxygen which is produced mostly by plants. Their influences become more manifest since no other live species can exist on earth without plants as they form the basic food staples too. Regulation of water cycle and oxygen production are the other roles of plants. The roles affect environment and climate. Plants are the main components of agricultural activities. Many countries benefit from these activities. Therefore, plants have impacts on political and economic situations and future of countries. Due to importance of plants and their roles, study of plants is essential in various fields. Consideration of their different applications leads to focus on details of them too. Automatic recognition of plants is a novel field to contribute other researches and future of studies. Moreover, plants can survive their life in different places and regions by means of adaptations. Therefore, adaptations are their special factors to help them in hard life situations. Weather condition is one of the parameters which affect plants life and their existence in one area. Recognition of plants in different weather conditions is a new window of research in the field. Only natural images are usable to consider weather conditions as new factors. Thus, it will be a generalized and useful system. In order to have a general system, distance from the camera to plants is considered as another factor. The other considered factor is change of light intensity in environment as it changes during the day. Adding these factors leads to a huge challenge to invent an accurate and secure system. Development of an efficient plant recognition system is essential and effective. One important component of plant is leaf which can be used to implement automatic systems for plant recognition without any human interface and interaction. Due to the nature of used images, characteristic investigation of plants is done. Leaves of plants are the first characteristics to select as trusty parts. Four different plant species are specified for the goal to classify them with an accurate system. The current paper is devoted to principal directions of the proposed methods and implemented system, image dataset, and results. The procedure of algorithm and classification is explained in details. First steps, feature detection and description of visual information, are outperformed by using Scale invariant feature transform (SIFT), HARRIS-SIFT, and FAST-SIFT methods. The accuracy of the implemented methods is computed. In addition to comparison, robustness and efficiency of results in different conditions are investigated and explained.

Keywords: SIFT combination, feature extraction, feature detection, natural images, natural plant recognition, HARRIS-SIFT, FAST-SIFT

Procedia PDF Downloads 243

Authors: Merve Begum Terzi, Orhan Arikan, Adnan Abaci, Mustafa Candemir

Abstract:

Acute myocardial infarction is a major cause of death in the world. Therefore, its fast and reliable diagnosis is a major clinical need. ECG is the most important diagnostic methodology which is used to make decisions about the management of the cardiovascular diseases. In patients with acute myocardial ischemia, temporary chest pains together with changes in ST segment and T wave of ECG occur shortly before the start of myocardial infarction. In this study, a technique which detects changes in ST/T sections of ECG is developed for the early diagnosis of acute myocardial ischemia. For this purpose, a database of real ECG recordings that contains a set of records from 75 patients presenting symptoms of chest pain who underwent elective percutaneous coronary intervention (PCI) is constituted. 12-lead ECG’s of the patients were recorded before and during the PCI procedure. Two ECG epochs, which are the pre-inflation ECG which is acquired before any catheter insertion and the occlusion ECG which is acquired during balloon inflation, are analyzed for each patient. By using pre-inflation and occlusion recordings, ECG features that are critical in the detection of acute myocardial ischemia are identified and the most discriminative features for the detection of acute myocardial ischemia are extracted. A classification technique based on support vector machine (SVM) approach operating with linear and radial basis function (RBF) kernels to detect ischemic events by using ST-T derived joint features from non-ischemic and ischemic states of the patients is developed. The dataset is randomly divided into training and testing sets and the training set is used to optimize SVM hyperparameters by using grid-search method and 10fold cross-validation. SVMs are designed specifically for each patient by tuning the kernel parameters in order to obtain the optimal classification performance results. As a result of implementing the developed classification technique to real ECG recordings, it is shown that the proposed technique provides highly reliable detections of the anomalies in ECG signals. Furthermore, to develop a detection technique that can be used in the absence of ECG recording obtained during healthy stage, the detection of acute myocardial ischemia based on ECG recordings of the patients obtained during ischemia is also investigated. For this purpose, a Gaussian mixture model (GMM) is used to represent the joint pdf of the most discriminating ECG features of myocardial ischemia. Then, a Neyman-Pearson type of approach is developed to provide detection of outliers that would correspond to acute myocardial ischemia. Neyman – Pearson decision strategy is used by computing the average log likelihood values of ECG segments and comparing them with a range of different threshold values. For different discrimination threshold values and number of ECG segments, probability of detection and probability of false alarm values are computed, and the corresponding ROC curves are obtained. The results indicate that increasing number of ECG segments provide higher performance for GMM based classification. Moreover, the comparison between the performances of SVM and GMM based classification showed that SVM provides higher classification performance results over ECG recordings of considerable number of patients.

Keywords: ECG classification, Gaussian mixture model, Neyman–Pearson approach, support vector machine

Procedia PDF Downloads 120

Authors: Zhao Gao, Eran Edirisinghe

Abstract:

The production of an accurate sketch of a suspect based on a verbal description obtained from a witness is an essential task for most criminal investigations. The criminal investigation system employs specifically trained professional artists to manually draw a facial image of the suspect according to the descriptions of an eyewitness for subsequent identification. Within the advancement of Deep Learning, Recurrent Neural Networks (RNN) have shown great promise in Natural Language Processing (NLP) tasks. Additionally, Generative Adversarial Networks (GAN) have also proven to be very effective in image generation. In this study, a trained GAN conditioned on textual features such as keywords automatically encoded from a verbal description of a human face using an RNN is used to generate photo-realistic facial images for criminal investigations. The intention of the proposed system is to map corresponding features into text generated from verbal descriptions. With this, it becomes possible to generate many reasonably accurate alternatives to which the witness can use to hopefully identify a suspect from. This reduces subjectivity in decision making both by the eyewitness and the artist while giving an opportunity for the witness to evaluate and reconsider decisions. Furthermore, the proposed approach benefits law enforcement agencies by reducing the time taken to physically draw each potential sketch, thus increasing response times and mitigating potentially malicious human intervention. With publically available 'CelebFaces Attributes Dataset' (CelebA) and additionally providing verbal description as training data, the proposed architecture is able to effectively produce facial structures from given text. Word Embeddings are learnt by applying the RNN architecture in order to perform semantic parsing, the output of which is fed into the GAN for synthesizing photo-realistic images. Rather than the grid search method, a metaheuristic search based on genetic algorithms is applied to evolve the network with the intent of achieving optimal hyperparameters in a fraction the time of a typical brute force approach. With the exception of the ‘CelebA’ training database, further novel test cases are supplied to the network for evaluation. Witness reports detailing criminals from Interpol or other law enforcement agencies are sampled on the network. Using the descriptions provided, samples are generated and compared with the ground truth images of a criminal in order to calculate the similarities. Two factors are used for performance evaluation: The Structural Similarity Index (SSIM) and the Peak Signal-to-Noise Ratio (PSNR). A high percentile output from this performance matrix should attribute to demonstrating the accuracy, in hope of proving that the proposed approach can be an effective tool for law enforcement agencies. The proposed approach to criminal facial image generation has potential to increase the ratio of criminal cases that can be ultimately resolved using eyewitness information gathering.

Keywords: RNN, GAN, NLP, facial composition, criminal investigation

Procedia PDF Downloads 136

Authors: Khachen Kongpakwattana, Nathorn Chaiyakunapruk

Abstract:

Background: Decision-analytic models for Alzheimer’s disease (AD) have been advanced to discrete-event simulation (DES), in which individual-level modelling of disease progression across continuous severity spectra and incorporation of key parameters such as treatment persistence into the model become feasible. This study aimed to apply the DES to perform a cost-effectiveness analysis of treatment for AD in Thailand. Methods: A dataset of Thai patients with AD, representing unique demographic and clinical characteristics, was bootstrapped to generate a baseline cohort of patients. Each patient was cloned and assigned to donepezil, galantamine, rivastigmine, memantine or no treatment. Throughout the simulation period, the model randomly assigned each patient to discrete events including hospital visits, treatment discontinuation and death. Correlated changes in cognitive and behavioral status over time were developed using patient-level data. Treatment effects were obtained from the most recent network meta-analysis. Treatment persistence, mortality and predictive equations for functional status, costs (Thai baht (THB) in 2017) and quality-adjusted life year (QALY) were derived from country-specific real-world data. The time horizon was 10 years, with a discount rate of 3% per annum. Cost-effectiveness was evaluated based on the willingness-to-pay (WTP) threshold of 160,000 THB/QALY gained (4,994 US$/QALY gained) in Thailand. Results: Under a societal perspective, only was the prescription of donepezil to AD patients with all disease-severity levels found to be cost-effective. Compared to untreated patients, although the patients receiving donepezil incurred a discounted additional costs of 2,161 THB, they experienced a discounted gain in QALY of 0.021, resulting in an incremental cost-effectiveness ratio (ICER) of 138,524 THB/QALY (4,062 US$/QALY). Besides, providing early treatment with donepezil to mild AD patients further reduced the ICER to 61,652 THB/QALY (1,808 US$/QALY). However, the dominance of donepezil appeared to wane when delayed treatment was given to a subgroup of moderate and severe AD patients [ICER: 284,388 THB/QALY (8,340 US$/QALY)]. Introduction of a treatment stopping rule when the Mini-Mental State Exam (MMSE) score goes below 10 to a mild AD cohort did not deteriorate the cost-effectiveness of donepezil at the current treatment persistence level. On the other hand, none of the AD medications was cost-effective when being considered under a healthcare perspective. Conclusions: The DES greatly enhances real-world representativeness of decision-analytic models for AD. Under a societal perspective, treatment with donepezil improves patient’s quality of life and is considered cost-effective when used to treat AD patients with all disease-severity levels in Thailand. The optimal treatment benefits are observed when donepezil is prescribed since the early course of AD. With healthcare budget constraints in Thailand, the implementation of donepezil coverage may be most likely possible when being considered starting with mild AD patients, along with the stopping rule introduced.

Keywords: Alzheimer's disease, cost-effectiveness analysis, discrete event simulation, health technology assessment

Procedia PDF Downloads 100

Authors: Maryam Fallahpoor, Biswajeet Pradhan

Abstract:

Background: Artificial intelligence (AI) serves as a valuable tool in mitigating the scarcity of human resources required for the evaluation and categorization of vast quantities of medical imaging data. When AI operates with optimal precision, it minimizes the demand for human interpretations and, thereby, reduces the burden on radiologists. Among various AI approaches, deep learning (DL) stands out as it obviates the need for feature extraction, a process that can impede classification, especially with intricate datasets. The advent of DL models has ushered in a new era in medical imaging, particularly in the context of COVID-19 detection. Traditional 2D imaging techniques exhibit limitations when applied to volumetric data, such as Computed Tomography (CT) scans. Medical images predominantly exist in one of two formats: neuroimaging informatics technology initiative (NIfTI) and digital imaging and communications in medicine (DICOM). Purpose: This study aims to employ DL for the classification of COVID-19-infected pulmonary patients and normal cases based on 3D CT scans while investigating the impact of image format. Material and Methods: The dataset used for model training and testing consisted of 1245 patients from IranMehr Hospital. All scans shared a matrix size of 512 × 512, although they exhibited varying slice numbers. Consequently, after loading the DICOM CT scans, image resampling and interpolation were performed to standardize the slice count. All images underwent cropping and resampling, resulting in uniform dimensions of 128 × 128 × 60. Resolution uniformity was achieved through resampling to 1 mm × 1 mm × 1 mm, and image intensities were confined to the range of (−1000, 400) Hounsfield units (HU). For classification purposes, positive pulmonary COVID-19 involvement was designated as 1, while normal images were assigned a value of 0. Subsequently, a U-net-based lung segmentation module was applied to obtain 3D segmented lung regions. The pre-processing stage included normalization, zero-centering, and shuffling. Four distinct 3D CNN models (ResNet152, ResNet50, DensNet169, and DensNet201) were employed in this study. Results: The findings revealed that the segmentation technique yielded superior results for DICOM images, which could be attributed to the potential loss of information during the conversion of original DICOM images to NIFTI format. Notably, ResNet152 and ResNet50 exhibited the highest accuracy at 90.0%, and the same models achieved the best F1 score at 87%. ResNet152 also secured the highest Area under the Curve (AUC) at 0.932. Regarding sensitivity and specificity, DensNet201 achieved the highest values at 93% and 96%, respectively. Conclusion: This study underscores the capacity of deep learning to classify COVID-19 pulmonary involvement using real 3D hospital data. The results underscore the significance of employing DICOM format 3D CT images alongside appropriate pre-processing techniques when training DL models for COVID-19 detection. This approach enhances the accuracy and reliability of diagnostic systems for COVID-19 detection.

Keywords: deep learning, COVID-19 detection, NIFTI format, DICOM format

Procedia PDF Downloads 45

Authors: Chevonne Reynolds, Res Altwegg, Andrew Balmford, Claire N. Spottiswoode

Abstract:

Modern agriculture has revolutionised the planet’s capacity to support humans, yet has simultaneously had a greater negative impact on biodiversity than any other human activity. Balancing the demand for food with the conservation of biodiversity is one of the most pressing issues of our time. Biodiversity-friendly farming (‘land sharing’), or alternatively, separation of conservation and production activities (‘land sparing’), are proposed as two strategies for mediating the trade-off between agriculture and biodiversity. However, there is much debate regarding the efficacy of each strategy, as this trade-off has typically been addressed by short term studies at fine spatial scales. These studies ignore processes that are relevant to biodiversity at larger scales, such as meta-population dynamics and landscape connectivity. Therefore, to better understand species response to agricultural land-use and provide evidence to underpin the planning of better production landscapes, we need to determine the merits of each strategy at larger scales. In South Africa, a remarkable citizen science project - the South African Bird Atlas Project 2 (SABAP2) – collates an extensive dataset describing the occurrence of birds at a 5-min by 5-min grid cell resolution. We use these data, along with fine-resolution data on agricultural land-use, to determine which strategy optimises the agriculture-biodiversity trade-off in a southern African context, and at a spatial scale never considered before. To empirically test this trade-off, we model bird species population density, derived for each 5-min grid cell by Royle-Nicols single-species occupancy modelling, against both the amount and configuration of different types of agricultural production in the same 5-min grid cell. In using both production amount and configuration, we can show not only how species population densities react to changes in yield, but also describe the production landscape patterns most conducive to conservation. Furthermore, the extent of both the SABAP2 and land-cover datasets allows us to test this trade-off across multiple regions to determine if bird populations respond in a consistent way and whether results can be extrapolated to other landscapes. We tested the land sparing/sharing trade-off for 281 bird species across three different biomes in South Africa. Overall, a higher proportion of species are classified as losers, and would benefit from land sparing. However, this proportion of loser-sparers is not consistent and varies across biomes and the different types of agricultural production. This is most likely because of differences in the intensity of agricultural land-use and the interactions between the differing types of natural vegetation and agriculture. Interestingly, we observe a higher number of species that benefit from agriculture than anticipated, suggesting that agriculture is a legitimate resource for certain bird species. Our results support those seen at smaller scales and across vastly different agricultural systems, that land sparing benefits the most species. However, our analysis suggests that land sparing needs to be implemented at spatial scales much larger than previously considered. Species persistence in agricultural landscapes will require the conservation of large tracts of land, and is an important consideration in developing countries, which are undergoing rapid agricultural development.

Keywords: agriculture, birds, land sharing, land sparing

Procedia PDF Downloads 183

Authors: Kaustav Mukherjee

Abstract:

In this paper we propose a Neural Style Transfer solution whereby we have created a Lightweight Separable Convolution Kernel Based GAN Architecture (SC-GAN) which will very useful for designing filter for Mobile Phone Cameras and also Edge Devices which will convert any image to its 2D ANIMATED COMIC STYLE Movies like HEMAN, SUPERMAN, JUNGLE-BOOK. This will help the 2D animation artist by relieving to create new characters from real life person's images without having to go for endless hours of manual labour drawing each and every pose of a cartoon. It can even be used to create scenes from real life images.This will reduce a huge amount of turn around time to make 2D animated movies and decrease cost in terms of manpower and time. In addition to that being extreme light-weight it can be used as camera filters capable of taking Comic Style Shots using mobile phone camera or edge device cameras like Raspberry Pi 4,NVIDIA Jetson NANO etc. Existing Methods like CartoonGAN with the model size close to 170 MB is too heavy weight for mobile phones and edge devices due to their scarcity in resources. Compared to the current state of the art our proposed method which has a total model size of 31 MB which clearly makes it ideal and ultra-efficient for designing of camera filters on low resource devices like mobile phones, tablets and edge devices running OS or RTOS. .Owing to use of high resolution input and usage of bigger convolution kernel size it produces richer resolution Comic-Style Pictures implementation with 6 times lesser number of parameters and with just 25 extra epoch trained on a dataset of less than 1000 which breaks the myth that all GAN need mammoth amount of data. Our network reduces the density of the Gan architecture by using Depthwise Separable Convolution which does the convolution operation on each of the RGB channels separately then we use a Point-Wise Convolution to bring back the network into required channel number using 1 by 1 kernel.This reduces the number of parameters substantially and makes it extreme light-weight and suitable for mobile phones and edge devices. The architecture mentioned in the present paper make use of Parameterised Batch Normalization Goodfellow etc al. (Deep Learning OPTIMIZATION FOR TRAINING DEEP MODELS page 320) which makes the network to use the advantage of Batch Norm for easier training while maintaining the non-linear feature capture by inducing the learnable parameters

Keywords: comic stylisation from camera image using GAN, creating 2D animated movie style custom stickers from images, depth-wise separable convolutional neural network for light-weight GAN architecture for EDGE devices, GAN architecture for 2D animated cartoonizing neural style, neural style transfer for edge, model distilation, perceptual loss

Procedia PDF Downloads 97

Authors: Jamie Lee Harder

Abstract:

In the context of rapidly changing social and economic circumstances in the developing world, this paper analyses access to public healthcare for poor people with disabilities in Cambodia. Like other countries of South East Asia, Cambodia is developing at rapid pace. The historical past of Cambodia, however, has set former social policy structures to zero. This past forces Cambodia and its citizens to implement new public health policies to align with the needs of social care, healthcare, and urban planning. In this context, the role of people with disabilities (PwDs) is crucial as new developments should and can take into consideration their specific needs from the beginning onwards. This paper is based on qualitative research with expert interviews and focus group discussions in Cambodia. During the field work it became clear that the identification tool for the poorest households (HHs) does not count disability as a financial risk to fall into poverty neither when becoming sick nor because of higher health expenditures and/or lower income because of the disability. The social risk group of poor PwDs faces several barriers in accessing public healthcare. The urbanization, the socio-economic health status, and opportunities for education; all influence social status and have an impact on the health situation of these individuals. Cambodia has various difficulties with providing access to people with disabilities, mostly due to barriers regarding finances, geography, quality of care, poor knowledge about their rights and negative social and cultural beliefs. Shortened budgets and the lack of prioritizations lead to the need for reorientation of local communities, international and national non-governmental organizations and social policy. The poorest HHs are identified with a questionnaire, the IDPoor program, for which the Ministry of Planning is responsible. The identified HHs receive an ‘Equity Card’ which provides access free of charge to public healthcare centers and hospitals among other benefits. The dataset usually does not include information about the disability status. Four focus group discussions (FGD) with 28 participants showed various barriers in accessing public healthcare. These barriers go far beyond a missing ramp to access the healthcare center. The contents of the FGDs were ratified and repeated during the expert interviews with the local Ministries, NGOs, international organizations and private persons working in the field. The participants of the FGDs faced and continue to face high discrimination, low capacity to work and earn an own income, dependency on others and less social competence in their lives. When discussing their health situation, we identified, a huge difference between those who are identified and hold an Equity Card and those who do not. Participants reported high costs without IDPoor identification, positive experiences when going to the health center in terms of attitude and treatment, low satisfaction with specific capacities for treatments, negative rumors, and discrimination with the consequence of fear to seek treatment in many cases. The problem of accessing public healthcare by risk groups can be adapted to situations in other countries.

Keywords: access, disability, health, inequality, Cambodia

Procedia PDF Downloads 125

Authors: Armin Nabaei, M. Omair Ahmad, M. N. S. Swamy

Abstract:

Many facial expression and emotion recognition methods in the traditional approaches of using LDA, PCA, and EBGM have been proposed. In recent years deep learning models have provided a unique platform addressing by automatically extracting the features for the detection of facial expression and emotions. However, deep networks require large training datasets to extract automatic features effectively. In this work, we propose an efficient emotion detection algorithm using face images when only small datasets are available for training. We design a deep network whose feature extraction capability is enhanced by utilizing several parallel modules between the input and output of the network, each focusing on the extraction of different types of coarse features with fined grained details to break the symmetry of produced information. In fact, we leverage long range dependencies, which is one of the main drawback of CNNs. We develop this work by introducing a Dynamic Soft-Margin SoftMax.The conventional SoftMax suffers from reaching to gold labels very soon, which take the model to over-fitting. Because it’s not able to determine adequately discriminant feature vectors for some variant class labels. We reduced the risk of over-fitting by using a dynamic shape of input tensor instead of static in SoftMax layer with specifying a desired Soft- Margin. In fact, it acts as a controller to how hard the model should work to push dissimilar embedding vectors apart. For the proposed Categorical Loss, by the objective of compacting the same class labels and separating different class labels in the normalized log domain.We select penalty for those predictions with high divergence from ground-truth labels.So, we shorten correct feature vectors and enlarge false prediction tensors, it means we assign more weights for those classes with conjunction to each other (namely, “hard labels to learn”). By doing this work, we constrain the model to generate more discriminate feature vectors for variant class labels. Finally, for the proposed optimizer, our focus is on solving weak convergence of Adam optimizer for a non-convex problem. Our noteworthy optimizer is working by an alternative updating gradient procedure with an exponential weighted moving average function for faster convergence and exploiting a weight decay method to help drastically reducing the learning rate near optima to reach the dominant local minimum. We demonstrate the superiority of our proposed work by surpassing the first rank of three widely used Facial Expression Recognition datasets with 93.30% on FER-2013, and 16% improvement compare to the first rank after 10 years, reaching to 90.73% on RAF-DB, and 100% k-fold average accuracy for CK+ dataset, and shown to provide a top performance to that provided by other networks, which require much larger training datasets.

Keywords: computer vision, facial expression recognition, machine learning, algorithms, depp learning, neural networks

Procedia PDF Downloads 51

Authors: Yan Zhang

Abstract:

Predictive maintenance is a technique to predict when an in-service machine will fail so that maintenance can be planned in advance. Analytics-driven predictive maintenance is gaining increasing attention in many industries such as manufacturing, utilities, aerospace, etc., along with the emerging demand of Internet of Things (IoT) applications and the maturity of technologies that support Big Data storage and processing. This study aims to build an end-to-end analytics solution that includes both real-time machine condition monitoring and machine learning based predictive analytics capabilities. The goal is to showcase a general predictive maintenance solution architecture, which suggests how the data generated from field machines can be collected, transmitted, stored, and analyzed. We use a publicly available aircraft engine run-to-failure dataset to illustrate the streaming analytics component and the batch failure prediction component. We outline the contributions of this study from four aspects. First, we compare the predictive maintenance problems from the view of the traditional reliability centered maintenance field, and from the view of the IoT applications. When evolving to the IoT era, predictive maintenance has shifted its focus from ensuring reliable machine operations to improve production/maintenance efficiency via any maintenance related tasks. It covers a variety of topics, including but not limited to: failure prediction, fault forecasting, failure detection and diagnosis, and recommendation of maintenance actions after failure. Second, we review the state-of-art technologies that enable a machine/device to transmit data all the way through the Cloud for storage and advanced analytics. These technologies vary drastically mainly based on the power source and functionality of the devices. For example, a consumer machine such as an elevator uses completely different data transmission protocols comparing to the sensor units in an environmental sensor network. The former may transfer data into the Cloud via WiFi directly. The latter usually uses radio communication inherent the network, and the data is stored in a staging data node before it can be transmitted into the Cloud when necessary. Third, we illustrate show to formulate a machine learning problem to predict machine fault/failures. By showing a step-by-step process of data labeling, feature engineering, model construction and evaluation, we share following experiences: (1) what are the specific data quality issues that have crucial impact on predictive maintenance use cases; (2) how to train and evaluate a model when training data contains inter-dependent records. Four, we review the tools available to build such a data pipeline that digests the data and produce insights. We show the tools we use including data injection, streaming data processing, machine learning model training, and the tool that coordinates/schedules different jobs. In addition, we show the visualization tool that creates rich data visualizations for both real-time insights and prediction results. To conclude, there are two key takeaways from this study. (1) It summarizes the landscape and challenges of predictive maintenance applications. (2) It takes an example in aerospace with publicly available data to illustrate each component in the proposed data pipeline and showcases how the solution can be deployed as a live demo.

Keywords: Internet of Things, machine learning, predictive maintenance, streaming data

Procedia PDF Downloads 358

Authors: Shanaka Gunasekara, Maleen Jayasuriya, Nalin Harischandra, Lilantha Samaranayake, Gamini Dissanayake

Abstract:

One serious facet of the worsening Human-Elephant conflict (HEC) in nations such as Sri Lanka involves elephant-train collisions. Endangered Asian elephants are maimed or killed during such accidents, which also often result in orphaned or disabled elephants, contributing to the phenomenon of lone elephants. These lone elephants are found to be more likely to attack villages and showcase aggressive behaviour, which further exacerbates the overall HEC. Furthermore, Railway Services incur significant financial losses and disruptions to services annually due to such accidents. Most elephant-train collisions occur due to a lack of adequate reaction time. This is due to the significant stopping distance requirements of trains, as the full braking force needs to be avoided to minimise the risk of derailment. Thus, poor driver visibility at sharp turns, nighttime operation, and poor weather conditions are often contributing factors to this problem. Initial investigations also indicate that most collisions occur in localised “hotspots” where elephant pathways/corridors intersect with railway tracks that border grazing land and watering holes. Taking these factors into consideration, this work proposes the leveraging of recent developments in Convolutional Neural Network (CNN) technology to detect elephants using an RGB/infrared capable camera around known hotspots along the railway track. The CNN was trained using a curated dataset of elephants collected on field visits to elephant sanctuaries and wildlife parks in Sri Lanka. With this vision-based detection system at its core, a prototype unit of an early warning system was designed and tested. This weatherised and waterproofed unit consists of a Reolink security camera which provides a wide field of view and range, an Nvidia Jetson Xavier computing unit, a rechargeable battery, and a solar panel for self-sufficient functioning. The prototype unit was designed to be a low-cost, low-power and small footprint device that can be mounted on infrastructures such as poles or trees. If an elephant is detected, an early warning message is communicated to the train driver using the GSM network. A mobile app for this purpose was also designed to ensure that the warning is clearly communicated. A centralized control station manages and communicates all information through the train station network to ensure coordination among important stakeholders. Initial results indicate that detection accuracy is sufficient under varying lighting situations, provided comprehensive training datasets that represent a wide range of challenging conditions are available. The overall hardware prototype was shown to be robust and reliable. We envision a network of such units may help contribute to reducing the problem of elephant-train collisions and has the potential to act as an important surveillance mechanism in dealing with the broader issue of human-elephant conflicts.

Keywords: computer vision, deep learning, human-elephant conflict, wildlife early warning technology

Procedia PDF Downloads 193

Authors: Xingyu Gao, Qiang Wu

Abstract:

Patents play a crucial role in protecting innovation and intellectual property. Early prediction of the impact of artificial intelligence (AI) patents helps researchers and companies allocate resources and make better decisions. Understanding the key factors that influence patent impact can assist researchers in gaining a better understanding of the evolution of AI technology and innovation trends. Therefore, identifying highly impactful patents early and providing support for them holds immeasurable value in accelerating technological progress, reducing research and development costs, and mitigating market positioning risks. Despite the extensive research on AI patents, accurately predicting their early impact remains a challenge. Traditional methods often consider only single factors or simple combinations, failing to comprehensively and accurately reflect the actual impact of patents. This paper utilized the artificial intelligence patent database from the United States Patent and Trademark Office and the Len.org patent retrieval platform to obtain specific information on 35,708 AI patents. Using six machine learning models, namely Multiple Linear Regression, Random Forest Regression, XGBoost Regression, LightGBM Regression, Support Vector Machine Regression, and K-Nearest Neighbors Regression, and using early indicators of patents as features, the paper comprehensively predicted the impact of patents from three aspects: technical, social, and economic. These aspects include the technical leadership of patents, the number of citations they receive, and their shared value. The SHAP (Shapley Additive exPlanations) metric was used to explain the predictions of the best model, quantifying the contribution of each feature to the model's predictions. The experimental results on the AI patent dataset indicate that, for all three target variables, LightGBM regression shows the best predictive performance. Specifically, patent novelty has the greatest impact on predicting the technical impact of patents and has a positive effect. Additionally, the number of owners, the number of backward citations, and the number of independent claims are all crucial and have a positive influence on predicting technical impact. In predicting the social impact of patents, the number of applicants is considered the most critical input variable, but it has a negative impact on social impact. At the same time, the number of independent claims, the number of owners, and the number of backward citations are also important predictive factors, and they have a positive effect on social impact. For predicting the economic impact of patents, the number of independent claims is considered the most important factor and has a positive impact on economic impact. The number of owners, the number of sibling countries or regions, and the size of the extended patent family also have a positive influence on economic impact. The study primarily relies on data from the United States Patent and Trademark Office for artificial intelligence patents. Future research could consider more comprehensive data sources, including artificial intelligence patent data, from a global perspective. While the study takes into account various factors, there may still be other important features not considered. In the future, factors such as patent implementation and market applications may be considered as they could have an impact on the influence of patents.

Keywords: patent influence, interpretable machine learning, predictive models, SHAP

Procedia PDF Downloads 18

Authors: Sudeshna Sengupta

Abstract:

Women’s work lives weave a complex fabric of myriad work relations and complex structures. Lives, when seen from the lens of work, is a saga of conjugated oppression by intertwined structures that are vertically and horizontally interwoven in a very complex manner. Women interact with multiple institutions through their work. The interactions and interplay of institutions shape their organization of work. They intersperse productive work with reproductive work, unpaid economic activities with unpaid care work, and all kinds of activities with leisure and self-care. The proposed paper intends to understand how women working as home-based workers in the National Capital Region (NCR) of India are organizing their everyday work, and how the organization of work is influenced by the interplay of structures. Situating itself in a multidisciplinary theoretical framework, this paper brings out how the gendering of work is playing out in the political, economic and social domain and shaping the work-life within the family, and in the paid workspace. The paper will use a primary data source, which is qualitative in nature. It will comprise 15 qualitative interviews of women home-based workers from the National Capital Region. The research uses a life history approach. The sampling was purposive using snowballing as a method. The dataset is part of the primary data (qualitative) collected for the ongoing Ph.D. work in Gender Studies at Ambedkar University Delhi. The home-based workers interviewed were in “non-factory” wage relations based on piece rates with flexible working hours. Their workplaces were their own homes with no spatial divide between living spaces and workspaces. Home-based workers were recognized as a group in the domain of labor economics in the 1980s. When menial work was cheaper than machine work, the capital owners preferred to outsource work as home-based work to women. These production spaces are fragmented and the identity of gender is created within labor processes to favor material accumulation. Both the employers and employees acknowledged the material gain of the capital owner when work was subcontracted to women at home. Simultaneously the market reinforced women’s reproductive role by conforming to patriarchal ideology. The contractors played an important role in implementing localized control on workers and also in finding workers for fragmented, gendered production processes. Their presence helped the employers in bringing together multiple forms of oppression that ranged from creating a structure to flout laws by creating shadow employers. It created an intertwined social and economic structure as well as a workspace where the line between productive and reproductive work gets blurred. The state invisibilized itself either by keeping the sector out of the domain of laws or by not implementing its own laws regulating working conditions or social security. It allowed the local hierarchy to function and define localized working conditions. The productive reproductive continuum reveals a labor control that influenced both the productive and reproductive work of women.

Keywords: informal sector, paid work, women workers, labor processes

Procedia PDF Downloads 141

Authors: Thilina Ranaweera, Enes Makalic, John L. Hopper, Adrian Bickerstaffe

Abstract:

Data are the primary asset of biomedical researchers, and the engine for both discovery and research translation. As the volume and complexity of research datasets increase, especially with new technologies such as large single nucleotide polymorphism (SNP) chips, so too does the requirement for software to manage, process and analyze the data. Researchers often need to execute complicated queries and conduct complex analyzes of large-scale datasets. Existing tools to analyze such data, and other types of high-dimensional data, unfortunately suffer from one or more major problems. They typically require a high level of computing expertise, are too simplistic (i.e., do not fit realistic models that allow for complex interactions), are limited by computing power, do not exploit the computing power of large-scale parallel architectures (e.g. supercomputers, GPU clusters etc.), or are limited in the types of analysis available, compounded by the fact that integrating new analysis methods is not straightforward. Solutions to these problems, such as those developed and implemented on parallel architectures, are currently available to only a relatively small portion of medical researchers with access and know-how. The past decade has seen a rapid expansion of data management systems for the medical domain. Much attention has been given to systems that manage phenotype datasets generated by medical studies. The introduction of heterogeneous genomic data for research subjects that reside in these systems has highlighted the need for substantial improvements in software architecture. To address this problem, we have developed SPARK, an enabling and translational system for medical research, leveraging existing high performance computing resources, and analysis techniques currently available or being developed. It builds these into The Ark, an open-source web-based system designed to manage medical data. SPARK provides a next-generation biomedical data management solution that is based upon a novel Micro-Service architecture and Big Data technologies. The system serves to demonstrate the applicability of Micro-Service architectures for the development of high performance computing applications. When applied to high-dimensional medical datasets such as genomic data, relational data management approaches with normalized data structures suffer from unfeasibly high execution times for basic operations such as insert (i.e. importing a GWAS dataset) and the queries that are typical of the genomics research domain. SPARK resolves these problems by incorporating non-relational NoSQL databases that have been driven by the emergence of Big Data. SPARK provides researchers across the world with user-friendly access to state-of-the-art data management and analysis tools while eliminating the need for high-level informatics and programming skills. The system will benefit health and medical research by eliminating the burden of large-scale data management, querying, cleaning, and analysis. SPARK represents a major advancement in genome research technologies, vastly reducing the burden of working with genomic datasets, and enabling cutting edge analysis approaches that have previously been out of reach for many medical researchers.

Keywords: biomedical research, genomics, information systems, software

Procedia PDF Downloads 237

Authors: Ksenia Volkova, Artur Petrosyan, Ignatii Dubyshkin, Alexei Ossadtchi

Abstract:

Brain-computer interfaces are a growing research field producing many implementations that find use in different fields and are used for research and practical purposes. Despite the popularity of the implementations using non-invasive neuroimaging methods, radical improvement of the state channel bandwidth and, thus, decoding accuracy is only possible by using invasive techniques. Electrocorticography (ECoG) is a minimally invasive neuroimaging method that provides highly informative brain activity signals, effective analysis of which requires the use of machine learning methods that are able to learn representations of complex patterns. Deep learning is a family of machine learning algorithms that allow learning representations of data with multiple levels of abstraction. This study explores the potential of deep learning approaches for ECoG processing, decoding movement intentions and the perception of proprioceptive information. To obtain synchronous recording of kinematic movement characteristics and corresponding electrical brain activity, a series of experiments were carried out, during which subjects performed finger movements at their own pace. Finger movements were recorded with a three-axis accelerometer, while ECoG was synchronously registered from the electrode strips that were implanted over the contralateral sensorimotor cortex. Then, multichannel ECoG signals were used to track finger movement trajectory characterized by accelerometer signal. This process was carried out both causally and non-causally, using different position of the ECoG data segment with respect to the accelerometer data stream. The recorded data was split into training and testing sets, containing continuous non-overlapping fragments of the multichannel ECoG. A deep convolutional neural network was implemented and trained, using 1-second segments of ECoG data from the training dataset as input. To assess the decoding accuracy, correlation coefficient r between the output of the model and the accelerometer readings was computed. After optimization of hyperparameters and training, the deep learning model allowed reasonably accurate causal decoding of finger movement with correlation coefficient r = 0.8. In contrast, the classical Wiener-filter like approach was able to achieve only 0.56 in the causal decoding mode. In the noncausal case, the traditional approach reached the accuracy of r = 0.69, which may be due to the presence of additional proprioceptive information. This result demonstrates that the deep neural network was able to effectively find a representation of the complex top-down information related to the actual movement rather than proprioception. The sensitivity analysis shows physiologically plausible pictures of the extent to which individual features (channel, wavelet subband) are utilized during the decoding procedure. In conclusion, the results of this study have demonstrated that a combination of a minimally invasive neuroimaging technique such as ECoG and advanced machine learning approaches allows decoding motion with high accuracy. Such setup provides means for control of devices with a large number of degrees of freedom as well as exploratory studies of the complex neural processes underlying movement execution.

Keywords: brain-computer interface, deep learning, ECoG, movement decoding, sensorimotor cortex

Procedia PDF Downloads 134

Authors: J. Ticona Chambi, E. A. De Almeida, C. A. Andrade Raymundo Gaiotto, A. M. Do Espírito Santo, L. Infantes, S. L. Cuffini

Abstract:

The solubility of active pharmaceutical ingredients (APIs) is challenging for the pharmaceutical industry. The new multicomponent crystalline forms as cocrystal and solvates present an opportunity to improve the solubility of APIs. Commonly, the procedure to obtain multicomponent crystalline forms of a drug starts by screening the drug molecule with the different coformers/solvents. However, it is necessary to develop methods to obtain multicomponent forms in an efficient way and with the least possible environmental impact. The Hansen Solubility Parameters (HSPs) is considered a tool to obtain theoretical knowledge of the solubility of the target compound in the chosen solvent. H-Bond Propensity (HBP), Molecular Complementarity (MC), Coordination Values (CV) are tools used for statistical prediction of cocrystals developed by the Cambridge Crystallographic Data Center (CCDC). The HSPs and the CCDC tools are based on inter- and intra-molecular interactions. The curcumin (Cur), target molecule, is commonly used as an anti‐inflammatory. The demethoxycurcumin (Demcur) and bisdemethoxycurcumin (Bisdcur) are natural analogues of Cur from turmeric. Those target molecules have differences in their solubilities. In this way, the work aimed to analyze and compare different tools for multicomponent forms prediction (solvates) of Cur, Demcur and Biscur. The HSP values were calculated for Cur, Demcur, and Biscur using the chemical group contribution methods and the statistical optimization from experimental data. The HSPmol software was used. From the HSPs of the target molecules and fifty solvents (listed in the HSP books), the relative energy difference (RED) was determined. The probability of the target molecules would be interacting with the solvent molecule was determined using the CCDC tools. A dataset of fifty molecules of different organic solvents was ranked for each prediction method and by a consensus ranking of different combinations: HSP, CV, HBP and MC values. Based on the prediction, 15 solvents were selected as Dimethyl Sulfoxide (DMSO), Tetrahydrofuran (THF), Acetonitrile (ACN), 1,4-Dioxane (DOX) and others. In a starting analysis, the slow evaporation technique from 50°C at room temperature and 4°C was used to obtain solvates. The single crystals were collected by using a Bruker D8 Venture diffractometer, detector Photon100. The data processing and crystal structure determination were performed using APEX3 and Olex2-1.5 software. According to the results, the HSPs (theoretical and optimized) and the Hansen solubility sphere for Cur, Demcur and Biscur were obtained. With respect to prediction analyses, a way to evaluate the predicting method was through the ranking and the consensus ranking position of solvates already reported in the literature. It was observed that the combination of HSP-CV obtained the best results when compared to the other methods. Furthermore, as a result of solvent selected, six new solvates, Cur-DOX, Cur-DMSO, Bicur-DOX, Bircur-THF, Demcur-DOX, Demcur-ACN and a new Biscur hydrate, were obtained. Crystal structures were determined for Cur-DOX, Biscur-DOX, Demcur-DOX and Bicur-Water. Moreover, the unit-cell parameter information for Cur-DMSO, Biscur-THF and Demcur-ACN were obtained. The preliminary results showed that the prediction method is showing a promising strategy to evaluate the possibility of forming multicomponent. It is currently working on obtaining multicomponent single crystals.

Keywords: curcumin, HSPs, prediction, solvates, solubility

Procedia PDF Downloads 36

Authors: Payam Mousavi, Andrew L. Stewart, Huiwen You, Aryeh F. G. Fayerman

Abstract:

We present a decision-support tool to assist an operator in the detection and tracking of a suspect vehicle traveling to an unknown target destination. Multiple data sources, such as traffic cameras, traffic information, weather, etc., are integrated and processed in real-time to infer a suspect’s intended destination chosen from a list of pre-determined high-value targets. Previously, we presented our work in the detection and tracking of vehicles using traffic and airborne cameras. Here, we focus on the fusion and processing of that information to predict a suspect’s behavior. The network of cameras is represented by a directional graph, where the edges correspond to direct road connections between the nodes and the edge weights are proportional to the average time it takes to travel from one node to another. For our experiments, we construct our graph based on the greater Los Angeles subset of the Caltrans’s “Performance Measurement System” (PeMS) dataset. We propose a Bayesian approach where a posterior probability for each target is continuously updated based on detections of the suspect in the live video feeds. Additionally, we introduce the concept of ‘soft interventions’, inspired by the field of Causal Inference. Soft interventions are herein defined as interventions that do not immediately interfere with the suspect’s movements; rather, a soft intervention may induce the suspect into making a new decision, ultimately making their intent more transparent. For example, a soft intervention could be temporarily closing a road a few blocks from the suspect’s current location, which may require the suspect to change their current course. The objective of these interventions is to gain the maximum amount of information about the suspect’s intent in the shortest possible time. Our system currently operates in a human-on-the-loop mode where at each step, a set of recommendations are presented to the operator to aid in decision-making. In principle, the system could operate autonomously, only prompting the operator for critical decisions, allowing the system to significantly scale up to larger areas and multiple suspects. Once the intended target is identified with sufficient confidence, the vehicle is reported to the authorities to take further action. Other recommendations include a selection of road closures, i.e., soft interventions, or to continue monitoring. We evaluate the performance of the proposed system using simulated scenarios where the suspect, starting at random locations, takes a noisy shortest path to their intended target. In all scenarios, the suspect’s intended target is unknown to our system. The decision thresholds are selected to maximize the chances of determining the suspect’s intended target in the minimum amount of time and with the smallest number of interventions. We conclude by discussing the limitations of our current approach to motivate a machine learning approach, based on reinforcement learning in order to relax some of the current limiting assumptions.

Keywords: autonomous surveillance, Bayesian reasoning, decision support, interventions, patterns of life, predictive analytics, predictive insights

Procedia PDF Downloads 93